JEV 对局回放 · 逐步看懂模型的思考

每一步都是一次 POST /v1/systemone 请求，请求体 = { state, model, questions }。 代码负责枚举全部合法落点并模拟后果，把数字词化成自然语言；Jev 不读数字、不做算术，只读词，返回带概率分布的判断。

① state.game —— 局面词化（第 1 步真实值）

② questions.placement —— 主问题：选哪个落点

Which placement of `game.current_piece` should the player choose? Each option describes the board after that placement.

优先级（权重全靠措辞顺序）：

1. 1. Clearing lines is good. Clearing more lines at once is better.
2. 2. Do not create holes. A placement with holes_created of none beats one that creates holes, unless the one with holes clears far more lines or the stack is dangerously high.
3. 3. Keep the stack low. Prefer a lower stack_height_after and a height_change that does not grow the stack.
4. 4. Keep the surface flat. Prefer surface_after of flat over slightly uneven, bumpy, or very jagged.
5. 5. One deep well is acceptable because the next I piece can fill it. Several deep wells are bad.
6. 6. When the stack is dangerously high, survival matters more than a clean surface.

每个候选落点一条 criteria，全部用词不用数，含显著事实（唯一/最高/最低等客观比较）。本局第 1 步共 17 个候选，示例 p0：

③ 投机问题（同请求随附，只进日志不影响落点）